BATsh-0.10

Changes for version 0.10 - 2026-08-01

No behaviour of the shell changes in this release: apart from the version numbers themselves, lib/ is untouched. Three CPAN Testers FAILs for 0.09 were defects of the tests rather than of the shell, and fixing them turned up a further group of assertions that could never fail, a launcher that could shadow its own installed module, and a distribution check that was enforcing something untrue.
TEST FIX (t/0019-tempfile-security.t): TS01 and TS03 assert that a temp file is created with open(O_CREAT|O_EXCL) and therefore never follows a symlink that some other process planted at the candidate path. That promise is POSIX. Win32 has no equivalent: CreateFile with CREATE_NEW resolves a reparse point and creates the target happily, so on a Windows smoker where symlink() happened to work (Windows 10 and later with developer mode) the two cases reported a difference in operating system semantics as a BATsh bug. Both are now skipped on the Windows family, with the reason printed, exactly as TS02 and TS04 already skipped their 0600 mode checks there. Reported for BATsh-0.09 on MSWin32 with perl 5.42.0.
TEST FIX (t/0015-cp932.t): CP12 writes to, and tests for, a file whose name contains the raw CP932 byte pair 0x83 0x5C. Not every environment can hold such a name: a Windows perl whose ANSI code page is UTF-8 rejects it with ENOENT before BATsh is reached. The case now probes the file system with plain perl first and skips, with the reason printed, when the name is unusable, instead of reporting a platform limit as a failure. Reported for BATsh-0.09 on MSWin32 with perl 5.42.0.
TEST FIX (t/0020-tilde-expansion.t): TE01 and TE02 compared the directory reached by "cd ~" against Cwd::realpath(). On Win32 that comparison is not portable -- realpath() and cwd() can disagree about separator direction, drive letter case and 8.3 short names -- and it produced a FAIL on a Windows smoker whose tilde expansion was in fact correct. Both now compare against the value cwd() reports after perl itself has chdir()ed to the same directory, which is the only reference that is guaranteed to be spelled identically. Reported for BATsh-0.09 on MSWin32 with perl 5.18.4.
TEST FIX (33 cases in t/0003, t/0006, t/0007, t/0009, t/0011, t/0020 and t/0032): an assertion written as
- _ok($out =~ /wanted/, 'name');
- evaluates the match in list context. A successful match with no capture group yields (1), so the call is _ok(1, 'name') and behaves; but a FAILED match yields the empty list, so the call collapses to _ok('name') -- a true value -- and the case passed no matter what the shell had printed. Every such assertion is now written
  - _ok(($out =~ /wanted/) ? 1 : 0, 'name');
- which restores scalar context. TE03 of t/0020 was the one case that had been silently passing over a real defect: it looks for a "cd: ... No such file or directory" diagnostic, which the shell writes to STDERR, while the helper captured STDOUT only. The helper now captures both streams and TE03 examines them together.
TEST FIX (t/9060-readme.t): the file defined a local _pm_version(), which INA_CPAN_Check has exported since 0.41, so every run of the suite printed "Subroutine _pm_version redefined at t/9060-readme.t line 47". The local copy is gone and the shared implementation is used. "pmake dist" is now warning free.
TEST FIX (t/0020, t/0026): the two cases that deliberately provoke a shell diagnostic ("cd: ~user: No such file or directory", "sh: ...: No such file or directory") let it through to the console, so a clean run looked like a broken one in the CPAN Testers output. Both helpers now capture STDERR as well.
CORRECTION: the 0.09 entry below states that writing "my (undef, ...)" to skip an element of a list assignment is a Perl 5.10 spelling. That is wrong, and this release withdraws the claim. Perl 5.005_03 already accepts the form: its op.c my() carries an explicit "else if (type == OP_UNDEF) return o;" branch, and perl 5.6.0 and 5.8.9 carry the same one, so the spelling has been legal across the whole range BATsh supports. Nothing has to change in the code -- the three declarations rewritten in 0.09 were correct before and are correct now -- but the rule that prompted the rewrite does, see the INA_CPAN_Check note below.
TEST FIX (t/0020-tilde-expansion.t): TE02 built its scratch directory with File::Temp, which is core only from 5.6.1. On 5.005_03 the "require File::Temp" inside its eval failed, the failure was swallowed, and the case reported success without having tested tilde expansion at all. The directory is now made with mkdir(), whose own failure is the atomicity guarantee, in the same spirit as the sysopen(O_CREAT|O_EXCL) used elsewhere in BATsh, and it is removed again at the end of the case.
FIX (bin/batsh.pl): the launcher added its sibling lib/ to @INC unconditionally. That is right when the script is run out of an unpacked distribution, but after installation into (say) /usr/local/bin it put /usr/local/lib at the *head* of @INC on every run, where a stale or unrelated BATsh.pm would shadow the copy that had just been installed. The directory is now added only when a MANIFEST sits beside it, which is what identifies an unpacked distribution and is never true of an installation prefix.
QA (t/9020-perl5compat.t): new check P13 rejects any module that entered core after 5.005_03, in lib/, t/, bin/ and eg/ alike. Wrapping such a load in eval does not satisfy it: that is precisely what let the File::Temp call above sit unnoticed, since on the oldest supported perl the eval simply failed and whatever depended on the module stopped being tested. Each scanned file now contributes 13 checks instead of 12.
QA (t/lib/INA_CPAN_Check.pm 0.41 -> 0.42): D4 used to reject "my (undef, ...)" as Perl 5.10 syntax. It is not, as described above, so the rule enforced a restriction that does not exist and could only cause correct code to be rewritten. The slot is reused for the post-5.005_03 core-module check that P13 also runs, sharing one list (@POST_5005_MODULES) so the two cannot drift apart. The number of checks per module is unchanged at six.
QA (pmake.bat 0.44 -> 0.45): runtime prerequisites were collected from lib/*.pm only, so a dependency that appeared solely in an installed bin/ script never reached META -- BATsh shipped without declaring FindBin, which bin/batsh.pl loads. The scan now covers bin/*.pl as well, and the "lib" pragma is dropped from the result alongside strict, warnings and vars.
QA (pmake.bat 0.45 -> 0.46): the prerequisite scan matched "use Module" only, so a module reached through "require Module" was invisible to META. BATsh loads Cwd and POSIX that way and declared neither. Both spellings are now collected, and they are not treated alike: "use Module" and a bare "require Module" are unconditional and become runtime requires, while a require inside eval is the pure Perl way of writing "use this if it is here" -- the caller has a fallback
- and becomes a recommendation instead. META.yml gains a "recommends" block (spec 1.4) and META.json a prereqs/runtime/recommends block (spec 2); Makefile.PL is unchanged, since a recommendation must not enter PREREQ_PM. BATsh now declares Cwd and POSIX under recommends, which is what the code actually means. The scan also skips POD and full-line comments, so prose that merely names a module can no longer invent a dependency.
QA (pmake.bat 0.46): "pmake dist" runs its own copy of the K1 rule over the sources listed in MANIFEST, and that copy did not say what INA_CPAN_Check::check_K says. It exempted a comma followed by $, a single quote or a double quote, on the reasoning that stripping a string literal leaves a quote character behind -- but f("a","b") really is missing the space, and so is f($x,$y), so the exemption was hiding exactly the lines the rule exists to find. The two are now spelled identically. Nothing in BATsh changes: the strict form finds no violation anywhere in the set "pmake dist" scans.
QA (t/lib/INA_CPAN_Check.pm 0.42 -> 0.43): style group K ran over lib/*.pm alone. Unlike D, E and G -- which demand a shape only a module is obliged to have -- K1, K2 and K3 ask for a space after a comma and for [ @array ] and { %hash } in place of \@array and \%hash, which are questions of spelling that apply to any Perl source. Restricting them to lib/ left most of a pure-Perl distribution unchecked, so K now runs over lib/, bin/ and t/ as listed in MANIFEST. Five lines needed respacing, none of which changes what the code does: t/0003 208, t/0005 40 and t/0009 273 open(STDOUT,'>&...') in the STDOUT-capture helpers, and t/9080 55 68 ok(0,"..."). With group K widened the suite runs 2071 tests.
QA (t/lib/INA_CPAN_Check.pm 0.43): new check B8. B1 to B7 tie $VERSION to META.yml, META.json, Makefile.PL and Changes, but the "Version 0.00" line printed under =head1 VERSION was verified by nothing -- G2 asks only that the heading exists. That line is what a reader sees on metacpan and it is a hand-edited copy of a number kept elsewhere, so it is precisely what goes stale unnoticed. B8 requires every shipped Perl source carrying the heading -- lib/*.pm and the bin/*.pl installed as EXE_FILES -- to state the distribution version under it. It is one check for the distribution, so check_B goes from six per module plus one to six per module plus two.
bin/batsh.pl: gains the =head1 VERSION section the five modules already carry, so the installed launcher documents which release it belongs to. B8 above keeps it from drifting.
Changes: this entry is dated 2026-08-01 JST, the day of the upload.